Robust skin-roughness estimation based on co-occurrence matrix
Automated segmentation of pigmented skin lesions (PSLs) from dermoscopy images is an important step for computeraided diagnosis of skin cancer. The segmentation task involves classifying each image pixel as either lesion or skin. It is challenging because lesion and skin can often have similar appearance. We present a novel exemplar-based algorithm for lesion segmentation which leverages the context provided by a global color model to retrieve annotated examples which are most similar to a given query image. Pixel labels are generated through a probabilistic voting rule and smoothed using a dermoscopy-specific spatial prior. We compare our method to three competing techniques using a large dataset of dermoscopy images with hand-segmented ground truth, We show that our exemplar-based approach yields significantly better segmentations and is computationally efficient.